Dans un contexte de données industrielles et numériques, le développement d'un outil sur mesure pour une tâche particulière est couteux par de nombreux aspects. À l'inverse, l'adaptation d'outils génériques l'est également en particularisation (personnalisation, adaptation, extension, …), pour les développeurs comme pour les utilisateurs finaux. Nos approches visent à considérer les différents niveaux d'interactions pour améliorer l'exploitation des données fournies ou générées en collaboration.Les définitions et problématiques liées aux données dépendent le plus souvent des domaines dans lesquelles elles sont traitées. Pour ce travail, nous avons opté pour une approche holistique considérant ensemble des perspectives différentes. Le résultat est une synthèse des concepts émergeant montrant les équivalences d'un domaine à l'autre. La première contribution consiste à améliorer le marquage collaboratif de documents. Deux améliorations sont proposées par notre outil Coviz. (1) L'étiquetage des ressources est propre à chaque utilisateur qui organise ses vocables par une poly-hiérarchie nominative. Chacun peut considérer les concepts des autres par une relation de partage. Le système fournit également du contenu connexe via un moissonnage des archives ouvertes. (2) L'outil applique le concept de facette des données à l'interface puis les combine avec une recherche par mot-clé. Ce dernier point est commun à tous les utilisateurs, le système considère chacune des actions individuelles comme celles d'un groupe.La contribution majeure, confidentielle, est un framework baptisé DIP pour Data Interaction and Presentation. Son but est d'augmenter la liberté d'expression de l'utilisateur sur l'interaction et l'accès aux données. Il diminue les contraintes machines et logicielles en adjoignant une nouvelle voix d'accès direct entre l'utilisateur et les données disponibles, ainsi que des points d'« articulation » génériques. D'un point de vue final, l'utilisateur gagne en expression de filtrage, en partage, en maintien de l'état de sa navigation, en automatisation de ses tâches courantes, etc.Il a été testé en condition réelle de stress, de ressources et d'utilisation avec le logiciel KeePlace. Ce dernier a d'ailleurs été l'initiateur de cette thèse. / In the context of numeric data, the software development costs entail a number of cost factors. In contrast, adapting generic tools has its own set of costs, requiring developer's integration and final user's adaptation. The aim of our approach is to consider the different points of interaction with the data to improve the exploitation of data, whether provided or generated from collaboration.The definitions and problems related to data are dependent upon the domain from which the data come and the treatment that have been applied to them. In this work we have opted for a holistic approach where we consider the range of angles. The result is a summary of the emergent concepts and domain equivalences.The first contribution consists of improving collaborative document mark-up. Two improvements are proposed by out tool – Coviz –. 1) Resource tagging which is unique to each user, who organises their own labels according to their personal poly-hierarchy. Each user may take into consideration other users approaches through sharing of tags. The system supplies additional context through a harvesting of documents in open archives. 2) The tool applies the concept of facets to the interface and then combines them to provide a search by keyword or characteristic selection. This point is shared by all users and the actions of an individual user impact the whole group.The major contribution, which is confidential, is a framework christened DIP for Data Interaction and Presentation. Its goal is to increase the freedom of expression of the user over the interaction and access to data. It reduces the hardware and software constrains by adding a new access point between the user and the raw data as well as generic pivots. From a final point of view the user gains in expression of filtering, in sharing, in state persistence of the navigator, in automation of day-to-day tasks, etc.DIP has been stress tested under real-life conditions of users and limited resources with the software KeePlace. Acknowledgement is given to KeePlace who initiated this thesis.

Treemaps are a popular space-filling visualization of hierarchical data that maps an attribute of a datum, or a data aggregate, to a proportional amount of area. Assuming a rectangular treemap consisting of nested rectangles (also called tiles), there are multiple possible valid tiling arrangements. A common criterion for optimization is aspect ratio. Nevertheless, treemaps usually consist of multiple rectangles, so the aspect ratios need be aggregated. The basic definition of aspect ratio (width divided by height) cannot be meaningfully aggregated. Given this, a definition of aspect ratio that does not differentiate height from width was suggested. This definition allows for meaningful aggregation, but only as long as there are no large differences in the data distribution, and the target aspect ratio is 1:1. Originally, a target aspect ratio of 1:1 was deemed to be axiomatically ideal. Currently, perceptual studies have found an aspect ratio of 1:1 to lead to the largest area estimation error. However, with any other target this definition of aspect ratio cannot be meaningfully aggregated. This thesis suggests a correction that can be applied to the current metric and would allow it to be meaningfully aggregated even when there are large value differences in the data. Furthermore, both the uncorrected and corrected metrics can be generalized for any target (i.e. targets other than 1:1). Another issue with current evaluation techniques is that algorithm fitness is evaluated through Monte Carlo trials. In this method, synthetic data is generated and then aggregated to generate a single final result. However, tiling algorithm performance is dependant on data distribution, so a single aggregateresult cannot generalize overall performance. The alternative suggested in this thesis is visual cluster analysis, which should hold more general predictive power.All of the above is put into practice with an experiment. In the experiment, a new family of tiling algorithms, based on criteria derived from the results of the perceptual tests in literature,is compared to the most popular tiling algorithm, Squarify. The results confirm that there are indeed vast but consistent value fluctuations for different normal distributions. At least for a target aspect ratio of 1.5, the new proposed algorithms are shown to perform better than Squarify for most use cases in terms of aspect ratio.